Lightgbm: [Feature] Python Bindings for LightGBM

[chivee]

Hi all, we have tentative plan on extending LightGBM to python users. please share us your opinions.

[ParadoxShmaradox]

I think this should be a great step forward with adoption, of course it should be also pretty easy to install and use.
These are the features I think you should consider for python support:

• install via pip
• scikit learn interface so people can use fit/predict and have a drop-in replacement for xgboost/GBM
• Loading/Saving models via pickle or dedicated methods

[yychenca]

Agree with ParadoxShmaradox. And for me earlystopping would be a great feature to have.

[chivee]

@ParadoxShmaradox @yychenca , we can break its procedure into two several phrases: 1. python bindings. 2 scikit-learn(or other popular framework) interface.

As for model (de)serialization, cross validation, and early stopping seems to be another popular features we should move on. I'll open issues for this features.

[geoHeil]

Want to add, that something like xgboosts scale_pos_weight to handle imbalanced classes, or even better an Implementation of _Example-Dependent Cost-Sensitive_ would be very interesting http://nbviewer.jupyter.org/github/albahnsen/CostSensitiveClassification/blob/master/doc/tutorials/tutorial_edcs_credit_scoring.ipynb

[mxbi]

I think we need a way to create the LightGBM binary format directly from a NumPy array instead of having to do an expensive write&read to CSV/SVMLight. For really big datasets (eg. Kaggle Bosch Competition) it can take an hour to read and write CSVs, so something like this (functionality similar to xgboost.DMatrix()) would be excellent.

[modkzs]

I think user-define loss and eval function is also important

[yychenca]

@chivee Thank you chivee. I had a chance to try LightGBM with a dataset of 180k rows and 30 features for a regression problem. The training completed in 8 seconds as compared to 52 secs by XGBoost using comparable parameters, 63 leaves vs 8 depths. And the accuracy of LIghtGBM (L1/MAE ) was even a bit better. Truly impressed!!

Noticed that early stopping has already been added and I'll probably give it another try soon.

Anyways, great job!

[ArdalanM]

A quick wrapper for LightGBM: https://github.com/ArdalanM/pyLightGBM
It still dump input to svm format before training on.

[chivee]

@ArdalanM , thank you ardalanM , that will be very helpful!

[chivee]

Hi guys, a simple Python binding using Ctype can be found at
https://github.com/Microsoft/LightGBM/blob/master/tests/c_api_test/test.py,

Any feedback to the bindings will be great.

[ternaus]

Existing python binding is great, but I would love to be able to replace existing algorithms in the pipeline and most likely they have scikit-learn interface.

At least on the level of:

• 
  
  1. fit
  
  
• 
  
  1. transform
  
  
• 
  
  1. fit_transform
  
  
• 
  
  1. predict
  
  
• 
  
  1. predict_proba
  
  

[guolinke]

refer to #94.
The basic class are finished. I think we can base on these classes to implement sklearn-liked interfaces.

[guolinke]

Hi, All,

the branch of python-package is merged, refer to https://github.com/Microsoft/LightGBM/tree/master/python-package
Welcome to have a try and provide feedback and issues.